Scene-aware robust fusion of side-ranging and forward-looking sonars for AUV perception in water-conveyance tunnels

Reliable acoustic perception is essential for autonomous underwater vehicle (AUV) inspection in water-conveyance tunnels. The main difficulty is that structural transitions, including width changes and main–branch junctions, alter sonar reliability and make genuine geometric variations difficult to distinguish from anomalous echoes. This paper proposes a scene-aware confidence-modulated robust sliding-window fusion method (SACM-RSWF) integrating four side-ranging sonars and a forward-looking sonar (FLS). FLS-derived structural descriptors infer straight, width-change, and junction probabilities, which regulate measurement confidence. Pseudo-Huber weighting and sparse outlier decomposition further suppress non-Gaussian acoustic errors. Numerical simulations are evaluated over 30 independent Monte Carlo runs, together with physical pool and Gazebo-based hardware-in-the-loop experiments. Relative to the RG-Huber baseline, SACM-RSWF reduces the mean customized normalized score by 13.0%, 23.6%, and 55.3% in the straight, width-change, and junction simulations, respectively. One-factor ablations further confirm the complementary contributions of different components. Experimental RMSE reductions of 22.6%–40.2% are achieved while maintaining an onboard processing rate of 11.45–11.81 Hz. The results demonstrate robust and real-time perception under structural ambiguity and unreliable acoustic observations.